Estimating Risk Levels and Epidemiology of Diabetic Retinopathy using Transfer Learning

Diabetic retinopathy is one among the visible microvascular effects of diabetes that may affect the retina of the human eye, and their images are now employed for manual disease evaluation and conclusion. Professional graders use feature-based evaluation criteria when using diagnostic methods for diabetic retinopathy (DR). Modern deep learning (DL) methods, on the other hand, are engrossed in end-to-end training and rely on image domain data labeling. The majority of these prognosis require a large training data to provide a straight categorization grade. This research is an attempt to accelerate the initial screening for DR on limited data in order to address the future needs of the diabetic community. In this study, we analyze the DL-based system trained on big data that can be further applied on smaller samples of retina image for screening of DR with the goal of supporting variable grades and track its progression. We developed and verified accurate classification models based on transfer learning to estimate the risk levels and epidemiology of DR using publicly available datasets, allowing for the prompt identification of DR. We applied data augmentation using Generative Adversarial Network to balance the dataset. For the grading of DR through multi-class (5-class) classification with augmented data, the precision varies from 0.40 to 0.97 on accessible datasets. The accuracy for detecting the presence of DR via binary classification ranges from 0.94 to 0.98, further adds to the model's efficacy in diagnosing diabetic retinopathy.